Groq

  Executive Briefing

Groq is a Mountain View–based AI infrastructure company that designs its own silicon specifically for LLM inference and operates one of the fastest publicly accessible inference APIs in the world. Founded in September 2016 by Jonathan Ross and Douglas Wightman — both former Google engineers — the company was built on a single conviction: that LLM inference has fundamentally different hardware requirements than training, demanding deterministic, ultra-low-latency throughput rather than raw GPU parallelism, and that the right way to meet those requirements is to build the chip from scratch. Ross had led the design of Google's original Tensor Processing Unit (TPU) before leaving to start Groq; Wightman came from Google X. The result of their work is the Language Processing Unit (LPU), a custom deterministic processor that eliminates the unpredictability of GPU scheduling and enables response times measured in fractions of a second at throughputs measured in hundreds to thousands of tokens per second.

Since launching GroqCloud as a public inference API in February 2024, Groq has assembled a three-tier platform: a developer-facing cloud API (OpenAI-compatible, with a free tier requiring no credit card), GroqRack on-premises hardware appliances for enterprise and government customers, and a technology-licensing program that culminated in a landmark deal. On 24 December 2025, Groq and Nvidia announced a non-exclusive inference-technology licensing agreement reported at approximately $20 billion — described as Nvidia's largest deal ever by value — in which Groq licensed its LPU architecture to Nvidia while co-founders Ross and president Sunny Madra transitioned to Nvidia.¹ Groq's independent operations continue under CEO Simon Edwards, who was appointed in late December 2025 following the Nvidia deal.² The deal's exact financial and legal structure remains under scrutiny, with the FTC examining it and a formal US Senate inquiry launched in March 2026 into its compliance with Hart-Scott-Rodino premerger notification rules.³

Groq's commercial footprint today spans more than 2 million developers, 12 data centers across the US, Canada, Saudi Arabia, and Europe (with Helsinki, Finland as the first European site, launched July 2025), and a series of strategic partnerships including Aramco Digital's world's-largest AI inference facility in Dammam, Saudi Arabia.⁴ The company has raised approximately $1.75 billion in disclosed equity rounds through its September 2025 Series E at a $6.9 billion post-money valuation, plus a reported $1.5 billion infrastructure commitment from the Kingdom of Saudi Arabia announced in February 2025 — figures that vary by source and should be treated as reported estimates.⁵⁶ Revenue has scaled rapidly, with 2024 reported at approximately $90 million and 2025 revenue variously reported or projected between $172 million and $500 million.⁷

The company occupies a distinctive niche in the inference infrastructure stack: unlike GPU cloud providers that use commodity Nvidia hardware, and unlike hyperscalers that bundle AI inference into broader cloud offerings, Groq owns its silicon, its compiler stack, and its deployment infrastructure end-to-end. Its benchmark speed numbers on open-weight models such as the Llama 3 family have made it the reference point for inference-speed comparisons across the industry.

  At a Glance

  Origins & Founding

Groq was founded in September 2016 by Jonathan Ross and Douglas Wightman, both alumni of Google. Ross had been a lead architect on the original Google Tensor Processing Unit (TPU) project — one of the most consequential custom-silicon efforts in computing history — before identifying that the inference workload was a fundamentally different problem from the matrix-multiply-heavy training workloads that TPUs were designed around.⁸ Inference, he argued, required deterministic latency, high memory bandwidth per chip, and the ability to respond to unpredictable real-time requests, none of which GPU architectures were designed to optimize. Wightman, coming from the more speculative and hardware-adjacent research environment of Google X, joined as co-founder. The pair operated in stealth for roughly three years, raising early capital from Social Capital and its founder Chamath Palihapitiya in 2017, before shipping their first silicon in 2019.⁹

The founding thesis was unusual in the AI infrastructure landscape of 2016: rather than building software that ran more efficiently on existing hardware, Ross and Wightman bet that you had to rebuild the chip to achieve the latency properties that next-generation AI applications would demand. That bet took years to mature — Groq was largely invisible to the developer ecosystem until the 2024 GroqCloud launch — but the public reception validated the approach dramatically, with the platform attracting more than 360,000 developers in its first 18 months.⁴

  History & Timeline

    2016–2019: Founding and first silicon

Groq was incorporated and began chip design work in 2016. Operating in stealth with early funding from Social Capital, the team built the first-generation Language Processing Unit on GlobalFoundries' 14nm process, internally referred to as the Tensor Streaming Processor (TSP). The first chips shipped in 2019, demonstrating exceptional linear-algebra throughput and, crucially, the deterministic latency property that distinguished the LPU from GPU inference. Early customers were primarily research institutions and defense-adjacent organizations who valued the predictable response-time guarantees.

    2020–2022: Enterprise hardware and the Maxeler acquisition

Through 2020 and 2021, Groq focused on selling its hardware as enterprise appliances and building out the compiler and toolchain ecosystem needed to compile arbitrary models onto the LPU architecture. The $300M Series C in April 2021, led by Tiger Global Management and D1 Capital Partners, brought the company to unicorn status at approximately $1 billion valuation and enabled significant manufacturing and go-to-market scale-up.¹⁰ In March 2022, Groq acquired Maxeler Technologies, a UK-based dataflow and high-performance computing company, absorbing their hardware and software engineering talent to accelerate next-generation chip development.

    2023–2024: Gen 2 silicon, GroqCloud launch, and Series D

In mid-2023 Groq announced that the second-generation LPU would be manufactured on Samsung's 4nm process at Samsung's Taylor, Texas foundry — a node that offers substantially higher transistor density and performance-per-watt than the 14nm Gen 1.¹¹ The Gen 2 chip delivers up to 750 TOPs (INT8), 188 TFLOPs (FP16 at 900 MHz), 230 MB of on-chip SRAM, and up to 80 TB/s of on-die memory bandwidth per card. The GroqCard accelerator is priced at approximately $19,948.

On February 2024, Groq launched GroqCloud as a public managed inference API, and the developer response was immediate and substantial.¹² The platform offered an OpenAI-compatible REST API, a free tier with 14,400 requests per day, and per-token pricing that positioned it as one of the most cost-competitive inference options for open-weight models. In March 2024, Groq also acquired Definitive Intelligence and signed a memorandum of understanding with Aramco Digital at the LEAP 2024 conference in Saudi Arabia. The $640M Series D closed on 5 August 2024, led by BlackRock Private Equity Partners, with Neuberger Berman, Cisco Investments, Samsung Catalyst Fund, and KDDI Open Innovation Fund also participating, at a $2.8B post-money valuation.¹³ In September 2024, Groq and Aramco Digital announced a partnership to build the world's largest AI inference data center in Dammam, Saudi Arabia; the facility came online in December 2024, reportedly in eight days.

    2025: Series E, Saudi infrastructure commitment, European expansion

The $750M Series E closed in September 2025, led by Disruptive, with returning investors BlackRock, Neuberger Berman, Samsung, Cisco, and D1 Capital Partners joined by Altimeter Capital, 1789 Capital, DTCP, and Infinitum — lifting the post-money valuation to $6.9 billion.¹⁴ In February 2025, the Kingdom of Saudi Arabia announced a $1.5 billion commitment to Groq AI infrastructure, building on the Aramco Digital partnership and establishing a GroqCloud regional hub through Humain, Saudi Arabia's AI company.⁵ In July 2025, Groq launched its first European data center in Helsinki, Finland, in partnership with Equinix, targeting European-based customers who need low-latency inference without data leaving the region.¹⁵

    December 2025 – Present: The Nvidia licensing deal and regulatory scrutiny

On 24 December 2025, Groq and Nvidia jointly announced a non-exclusive inference-technology licensing agreement reported at approximately $20 billion in value — the largest such deal in Nvidia's reported history.¹² Nvidia licensed Groq's LPU architecture and related inference technology on a non-exclusive basis. As part of the arrangement, co-founders Jonathan Ross and president Sunny Madra transitioned to Nvidia, while Simon Edwards, Groq's head of finance, was appointed CEO of the continuing independent entity.¹⁶ GroqCloud continued operations uninterrupted.

The deal's structure attracted immediate regulatory attention. In January 2026, the FTC began examining it as a potential "merger in disguise" — a structure that observers compared to Microsoft's 2024 arrangement with Inflection AI and Amazon's 2024 deal with Adept AI. On 20 March 2026, US Senators Elizabeth Warren and Richard Blumenthal launched a formal Senate inquiry into whether the agreement violated Hart-Scott-Rodino premerger notification requirements.³ As of June 2026, no formal enforcement action has been announced, but the inquiry remains open. Nvidia, for its part, publicly characterized the arrangement as a straightforward technology license, not an acquisition.¹⁷

  What They Offer — Products & Platform

Groq's platform spans three integrated tiers, each targeting a distinct buyer profile.

GroqCloud is the company's managed public inference API, launched February 2024. It provides an OpenAI-compatible REST endpoint that allows applications to migrate from other providers with minimal code changes. The free tier allows 14,400 requests per day across all models with a 30-request-per-minute rate limit, no credit card required — the most generous free tier among dedicated inference API providers. Paid tiers are consumption-priced per token. The platform also offers a Batch API with a 50% discount on on-demand pricing for non-real-time workloads, and prompt caching that applies an additional 50% discount on cached input tokens, stackable to approximately 25% of base on-demand pricing. Audio transcription (Whisper) and text-to-speech (Orpheus TTS models from Canopy Labs) are available alongside text generation endpoints.

GroqRack is Groq's on-premises enterprise hardware offering: integrated rack-scale systems containing 64 to 576 or more LPUs, sold as appliances to government agencies, research institutions, defense contractors, and private-cloud operators who require data sovereignty or cannot route sensitive workloads through a public API. GroqRack installations are typically reported to exceed $1 million per installation, and Groq works with federal resellers including Carahsoft to reach US government buyers.

Technology licensing became the third leg of the platform with the December 2025 Nvidia agreement, in which Groq licensed its LPU architecture on a non-exclusive basis. The exact financial structure — described by some sources as cash payments across three installments through 2026 — is reported but unverified; the $20 billion figure and payment schedule have not been officially confirmed by either party and should be treated as reported estimates.¹¹⁷

  Technology & Infrastructure

    The Language Processing Unit (LPU)

The core of Groq's technology is the Language Processing Unit, internally architected as a Tensor Streaming Processor (TSP). The LPU's design philosophy departs radically from GPU architecture in several interconnected ways, and understanding these departures explains both the LPU's performance advantages and its engineering constraints.

A conventional GPU achieves high throughput by running thousands of threads in parallel, using branch predictors and speculative execution to keep functional units busy. The GPU's memory hierarchy relies on High Bandwidth Memory (HBM) — fast stacked DRAM located near the die — to feed the compute units. For inference, this creates two problems: HBM latency is non-deterministic (influenced by memory scheduling, thermal throttling, and competing workloads), and the GPU's general-purpose threading model carries overhead that manifests as irreducible jitter in time-to-first-token.

The LPU eliminates both problems through a single-core deterministic design. There are no branch predictors and no speculative execution. Instead of HBM, the LPU uses compiler-managed flat SRAM as its primary weight storage — hundreds of megabytes of on-chip SRAM per chip, which is approximately 20 times faster to access than HBM.⁸ All data paths are fixed at compile time by Groq's ahead-of-time compiler: the compiler schedules every operation into a static execution timeline before the chip runs, so at runtime the processor simply executes the pre-computed schedule with no scheduling overhead. This static schedule is the source of the LPU's determinism: given the same request, the response-time bound is guaranteed by the compiled schedule rather than probabilistic under congestion.

Multiple LPUs connect via direct chip-to-chip plesiosynchronous interconnect, allowing them to act as a single logical processing unit for large models whose weights exceed the SRAM capacity of a single chip. Unlike GPU NVLink topologies that use HBM as a staging buffer between chips, LPU inter-chip transfers operate at on-chip SRAM speeds. The design is also air-cooled by default — a meaningful operational advantage for on-premises deployments where liquid cooling infrastructure is absent or undesirable.

Gen 1 LPU (GlobalFoundries 14nm process, shipped 2019): first-generation production silicon, demonstrated the viability of the static-scheduling approach.

Gen 2 LPU (Samsung 4nm process, announced 2023, manufactured at Samsung's Taylor, Texas foundry): the current generation underlying GroqCloud. Key specifications:

Gen 2 deployment is expected to significantly increase throughput per rack compared to Gen 1; full mass-production status and exact deployment timeline for Gen 2 in all GroqCloud regions have not been fully confirmed in public sources and should be treated as reported estimates.¹¹

    Data Centers and Global Footprint

As of mid-2026, Groq operates 12 data-center facilities:

• United States — Multiple sites across colocation providers including Equinix, TierPoint, and DataBank
• Canada — Bell Canada deployment
• Saudi Arabia — Dammam facility, built in partnership with Aramco Digital and Humain, one of the world's largest AI inference facilities, targeting hundreds of billions of tokens per day
• Europe — Helsinki, Finland (launched July 2025, Groq's first European site, in partnership with Equinix)¹⁵

Groq has announced plans to add 12 or more additional data-center sites in 2026, including first expansion into Asia. Aggregate capacity is reported at 20 million tokens per second across all regions as of mid-2025.⁴

    Compiler and Software Stack

The LPU's static-scheduling architecture requires that every model be compiled before it can be served. Groq ships a full ahead-of-time compiler toolchain that ingests standard model formats (ONNX, PyTorch) and produces a static execution schedule for the target number of LPU chips. This compilation step is opaque to end users of GroqCloud — they call the API as they would any inference endpoint — but it means that model support requires explicit compilation work by Groq's engineering team, which constrains the breadth of the model catalog compared to GPU-based providers that can run arbitrary CUDA-compatible models.

  Model Catalog & Performance

GroqCloud hosts open-weight models from leading labs, running on Groq's own LPU silicon. Groq does not train the models it serves. The catalog as of mid-2026 includes:

Meta AI Llama family

• Llama 3.3 70B Versatile — flagship instruction-tuned 70B model; 1,660+ tokens/sec with Speculative Decoding variant
• Llama 3.1 8B Instant — fastest available; 1,345 tokens/sec typical
• Llama 3.1 70B Versatile
• Llama 3.1 405B — available via partnership

OpenAI open-weight models

• openai/gpt-oss-120b — $0.15/M input, $0.60/M output
• openai/gpt-oss-safeguard-20b — content moderation

Qwen family

• Qwen QwQ 32B — reasoning model
• Qwen 2.5 32B
• Qwen 2.5 Coder 32B

DeepSeek distillations

• DeepSeek R1 Distill Qwen 32B
• DeepSeek R1 Distill Llama 70B

Mistral

• Mistral Saba 24B

Audio / Multimodal

• Whisper Large v3 — speech-to-text
• Orpheus TTS (Canopy Labs) — text-to-speech

  Pricing & Performance Position

GroqCloud's pricing is structured around pay-per-token consumption, with the free tier remaining the most accessible entry point of any major inference provider.

    Token Pricing (as of mid-2026)

Batch API: 50% discount on all models. Prompt caching: 50% discount on cached input tokens. Both discounts are stackable, yielding approximately 25% of on-demand pricing for heavily repeated, batch-processed requests.

Free tier: 14,400 requests/day across all models; 30 req/min rate limit; no credit card required.

    Throughput and Latency

These figures represent publicly reported benchmarks; actual throughput varies by request length, concurrency, and data-center region.

  People & Leadership

Jonathan Ross co-founded Groq in 2016 after leading the design of Google's original TPU. He served as CEO from founding through December 2025, was named to TIME magazine's 100 Most Influential People in AI in 2024, and joined Nvidia as part of the December 2025 licensing agreement.¹ Douglas Wightman co-founded Groq alongside Ross, bringing hardware and systems expertise from Google X; he remains with Groq as co-founder. Sunny Madra served as President until December 2025 and also transitioned to Nvidia with Ross.¹⁶

Simon Edwards, formerly Groq's head of finance, was appointed CEO of Groq following the Nvidia deal in late December 2025.¹⁶ Edwards has limited public profile relative to his predecessor, and an official executive biography from Groq had not been widely published as of mid-2026; his appointment is confirmed by press reports.

Early backers include Chamath Palihapitiya of Social Capital, who led the 2017 seed round.⁹

  Funding, Ownership & Business

Groq has raised equity across five disclosed rounds, plus a large sovereign infrastructure commitment:

Total disclosed equity raised is approximately $1.75 billion through the Series E.¹³¹⁴ Sources citing a higher figure of up to $3.35 billion may be including the $1.5 billion Saudi infrastructure commitment as equity-equivalent — a characterization that has not been officially confirmed. The Saudi commitment is described as an infrastructure expansion commitment rather than a standard equity round.⁵

The Nvidia licensing agreement, reported at approximately $20 billion in value, constitutes the largest single transaction in Groq's history if the reported figures are accurate. Neither company has officially confirmed the $20 billion figure or the payment schedule; the valuation and structure are reported estimates from media sources, and Nvidia publicly stated that reports characterizing the deal as an "acquisition" were inaccurate.¹⁷

Groq's business model operates across three revenue streams: (1) GroqCloud API consumption revenue (pay-per-token, batch, and caching); (2) GroqRack enterprise hardware appliance sales; and (3) technology licensing fees. Revenue for 2024 is reported at approximately $90 million. Revenue for 2025 is variously reported at $172.5 million (Latka database) or projected at $500 million (figure reportedly provided to investors) — the gap reflects the difference between actual measured revenue and a forward projection; the $500 million figure's achievement is unconfirmed.⁷ Company-reported revenue projections for 2026 and 2027 are $1.2 billion and $1.9 billion respectively, reflecting anticipated growth from the Saudi hub, European expansion, and broader enterprise hardware sales.

  Customers & Partnerships

Aramco Digital / Humain (Saudi Arabia) — Groq's most prominent infrastructure partnership. The Dammam facility, built jointly with Aramco Digital and scaled through the Saudi sovereign commitment via Humain, is targeting capacity to handle hundreds of billions of tokens per day, positioning Saudi Arabia as a regional hub for AI inference in the Middle East.⁴

Bell Canada — GroqRack enterprise deployment, one of Groq's flagship Canadian infrastructure partners.

IBM — Enterprise inference partner for business customers seeking high-throughput, low-latency model serving.

Carahsoft — Authorized federal reseller for GroqRack and GroqCloud in US government markets, enabling Groq to reach defense, civilian agency, and national lab customers.

Equinix — Infrastructure partnership for European and US data-center colocation, enabling the Helsinki launch and low-latency European deployments.¹⁵

KDDI (Japan) — Strategic investor (Series D) with potential for carrier-grade inference deployment across Japanese telecommunications infrastructure.

Cisco — Strategic investor (Series D and E) and enterprise integration partner, enabling Groq services to be bundled into Cisco's enterprise networking and security product lines.

Nvidia — Technology licensing partner as of December 2025; the non-exclusive agreement licenses Groq's LPU architecture to Nvidia without transferring Groq's ongoing business or GroqCloud operations.¹

Groq reports a developer community of more than 2 million users, with over 360,000 developers onboarded in the first 18 months following the GroqCloud launch. Sacra equity research has reported that approximately 75% of Fortune 100 companies maintain GroqCloud accounts, though this figure has not been independently verified.⁷

  Competitive Position

Groq competes in the custom-silicon inference segment against two primary peers: Cerebras Systems, which builds wafer-scale processors, and SambaNova Systems, which uses reconfigurable dataflow RDUs.

In head-to-head benchmarks on the Llama 3.1 8B model, independent analysis has placed Cerebras first (approximately 1,837 tokens/sec), SambaNova second (approximately 988 tokens/sec), and Groq third (approximately 750 tokens/sec on standard configuration).¹⁸ On 70B parameter models, SambaNova typically leads. Groq's differentiators in this comparison are its deterministic latency guarantee, the breadth of its public model catalog, and the maturity of its developer-facing GroqCloud API — the most widely adopted public interface of the three.

Against GPU-based inference providers — Together AI, Fireworks AI, Replicate, AWS Bedrock, Azure AI Foundry, and others — Groq positions on a 10x or greater speed advantage and sub-0.8-second time-to-first-token on supported models. GPU providers can serve a broader range of models including fine-tuned and custom-weight variants; Groq's static compilation requirement means the catalog is narrower but each supported model runs faster and with more predictable latency.

The Nvidia licensing deal introduces a new strategic dimension. By embedding Groq's LPU architecture into Nvidia's future product roadmap, the deal validates the deterministic-inference approach at the industry's highest level while potentially reducing Groq's ability to compete on silicon differentiation if Nvidia ships LPU-derived capabilities in mainstream products. Conversely, licensing revenue at the reported scale transforms Groq's balance sheet independent of inference API revenue growth.

  Notable Events

December 24, 2025 — Nvidia licensing agreement: Groq and Nvidia announce a non-exclusive inference-technology licensing agreement reported at approximately $20 billion in value. Jonathan Ross and Sunny Madra join Nvidia. Simon Edwards becomes CEO of the continuing Groq entity. The deal is Nvidia's largest reported deal by value.¹

January 2026 — FTC examination: The US Federal Trade Commission begins examining the Nvidia-Groq arrangement under the framework of "merger in disguise" structures, following a pattern established by scrutiny of the Microsoft-Inflection and Amazon-Adept deals from 2024.³

March 20, 2026 — Senate inquiry: US Senators Elizabeth Warren and Richard Blumenthal launch a formal Senate inquiry into whether the Nvidia-Groq deal violates Hart-Scott-Rodino premerger notification requirements, which would require parties to report and observe a waiting period before closing transactions above certain thresholds.³ As of June 2026, no formal enforcement action has been announced; the inquiry remains open.

December 2024 — Dammam data center online: The Aramco Digital partnership facility in Dammam, Saudi Arabia comes online, reportedly constructed in eight days — a deployment speed demonstration that Groq uses as a reference for its GroqRack on-premises installation proposition.⁴

July 2025 — Helsinki launch: Groq opens its first European data center in Helsinki, Finland, in partnership with Equinix, enabling European customers to use GroqCloud with data remaining in the EU/EEA region.¹⁵

  Outlook & Roadmap

Groq under CEO Simon Edwards is pursuing several concurrent growth vectors. Infrastructure expansion remains the primary capital deployment priority, with 12 or more additional data-center sites announced for 2026, including first deployments in Asia. The Saudi Arabia GroqCloud hub is scaling toward hundreds of billions of tokens per day, supported by the $1.5 billion sovereign commitment. A FedRAMP certification roadmap is in progress to allow GroqCloud to serve US federal civilian and defense customers through the company's own platform rather than exclusively through Carahsoft reseller channels.

The Gen 2 LPU on Samsung's 4nm process, deployed in GroqCloud regions through 2025 and 2026, is expected to significantly increase throughput per rack over Gen 1 silicon. This capacity expansion should enable Groq to serve larger concurrent workloads and bring per-token cost down as the platform scales.

Company-reported revenue projections for 2026 ($1.2 billion) and 2027 ($1.9 billion) reflect management's expectations for compounding growth from the API platform, enterprise hardware, and the Saudi infrastructure hub; these are forward projections and should be treated as estimates rather than realized figures.⁷

The dominant uncertainty for Groq's near-term trajectory is regulatory. If the FTC or DOJ ultimately reclassifies the Nvidia licensing agreement as a reportable acquisition subject to merger review, Groq could face orders to unwind the license, divest related technology transfers, or modify the acqui-hire arrangements involving Ross and Madra. The parallel Senate inquiry focuses on procedural compliance but could intensify pressure for formal review. No enforcement timeline has been established as of June 2026.

More broadly, Groq's positioning as a sovereign AI infrastructure provider — selling LPU capacity to government entities, national telecom carriers, and sovereign wealth-backed AI programs — appears to be the fastest-growing strategic wedge beyond the developer-focused GroqCloud API. The combination of deterministic latency, air-cooled hardware suitable for diverse facility types, and the credibility conferred by the Nvidia licensing deal positions Groq as infrastructure of choice for national AI programs that cannot or will not depend exclusively on US hyperscaler clouds.

---

  References

1. Groq and Nvidia — Non-Exclusive Inference Technology Licensing Agreement (Groq Newsroom)
2. Nvidia reportedly confirms non-acquisition structure (TrendForce, Dec 2025)
3. CNBC — Nvidia acquiring AI chip startup Groq for ~$20B (Dec 2025)
4. Groq — Aramco Digital Data Center Partnership
5. Saudi Arabia $1.5B Groq Infrastructure Commitment (Groq Newsroom)
6. Groq — Series E $750M raise (Groq Newsroom)
7. Groq revenue and customer figures — Sacra equity research
8. Groq LPU Architecture
9. Groq — Wikipedia
10. TechCrunch — Groq Series C (Apr 2021, archived)
11. Samsung Taylor, Texas foundry — 4nm production (industry reporting)
12. GroqCloud public launch — Feb 2024 (Groq)
13. Groq — Series D $640M raise (Groq Newsroom)
14. Groq — Series E announcement
15. Groq — Helsinki, Finland European data center launch
16. Data Center Dynamics — Nvidia to license tech from Groq, hire its leadership
17. TrendForce — Nvidia reportedly denies $20B Groq acquisition (Dec 2025)
18. Cerebras vs SambaNova vs Groq AI chips (Intuition Labs)

  References

1. Groq and Nvidia licensing announcement — Groq Newsroom; CNBC reporting — CNBC Dec 2025. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

2. Simon Edwards appointment as CEO — Data Center Dynamics; see also TechCrunch reporting — TechCrunch Dec 2025. ↩ ↩²

3. FTC examination and Senate inquiry — reported by CNBC and industry press; no official FTC enforcement document publicly available as of June 2026. ↩ ↩² ↩³ ↩⁴

4. Aramco Digital partnership and developer figures — Groq Newsroom. ↩ ↩² ↩³ ↩⁴ ↩⁵

5. Saudi Arabia $1.5B commitment — Groq Newsroom. ↩ ↩² ↩³

6. Series E valuation — Groq Newsroom. ↩

7. Revenue and customer figures are reported/estimated; the $500M 2025 projection is a figure Groq reportedly provided to investors, not a confirmed result — Sacra. ↩ ↩² ↩³ ↩⁴

8. LPU architecture overview — Groq; founding context — Wikipedia: Groq. ↩ ↩²

9. Seed round and Social Capital — Wikipedia: Groq. ↩ ↩²

10. Series C — Tiger Global and D1 Capital — Wikipedia: Groq. ↩

11. Gen 2 LPU on Samsung 4nm — deployment status partially confirmed; full production timeline is reported/estimated — Data Center Dynamics. ↩ ↩²

12. GroqCloud launch February 2024 — Groq. ↩

13. Series D — Groq Newsroom. ↩ ↩²

14. Series E — Groq Newsroom. ↩ ↩²

15. Helsinki European data center — Groq Newsroom; Equinix blog. ↩ ↩² ↩³ ↩⁴

16. Ross, Madra, and Edwards transition — Data Center Dynamics. ↩ ↩² ↩³

17. Nvidia denial of acquisition characterization — TrendForce. ↩ ↩² ↩³

18. Custom silicon inference benchmark comparison — Intuition Labs. ↩